Purpose:
Build a product that helps doctors quickly identify cases of pneumonia in children.
Output:
- Created a data labeling job using Appen's platform.
- Developed a proposal file, which is a writeup that details my design considerations and strategies for quality assurance.
Purpose: Study how training data impact models.
Description: I trained four different models using four variants of the pneumonia dataset classified into two classes: "normal" and "pneumonia", with Google's AutoML Vision platform. The following sections describe the steps we took to create each model.
Steps:
- Created a binary classifier to detect pneumonia using chest x-rays (100 images from the “normal” class and 100 images from the “pneumonia” class)
- Created an unbalanced binary classifier (100 images from the “normal” class, and add 200 more "pneumonia" class images)
- Created a binary classifier with dirty data (started with the original dataset of 100 "normal" and 100 "pneumonia" images. Then switched the labels of 30 images in each class)
- Created a three-class model with the classes “normal”, “bacterial pneumonia”, and “viral pneumonia” (added 100 "normal" images, 100 "bacterial pneumonia" images, and 100 "viral pneumonia" images)
Purpose:
The project purpose is to detect if a child is affected by pneumonia or not by using a chest x-ray images.
The aim is to elaborate a project proposal while accounting for:
- Data Labeling Approach
- Test Questions & Quality Assurance
- Limitations & Improvements